The Learning Crisis of Developing Country Elites: Lessons from PISA-D

Abstract

How much of the learning crisis can be addressed through “inclusion”—the equalization of grade attainment and learning outcomes across groups (e.g., girls/boys, rural/urban, poor/rich)—and how much of the learning crisis requires improvement in the country's system of basic education to improve learning outcomes across the board? This study uses the data from the seven countries who participated in Programme for International Student Assessment (PISA) for Development (PISA-D) to show that for most countries and subjects the average learning outcome for the advantaged (male, urban, native-born, speakers of assessment language), and elite (95th percentile in PISA measured socio-economic status) students was below the Sustainable Development Goal (SDG) “minimum learning level” target of PISA level 2. Even if every child in these countries were fully “included”—had the same distribution of learning outcomes as the advantaged, SES elite, public school children—80 percent of all children would still fall short of proposed global minimum levels of learning.

1. Introduction

This study examines Programme for International Student Assessment (PISA) for Development (PISA-D) data for seven countries, which have learning profiles typical of developing countries, and establishes three important facts:

1. First, in the sampled universe of students, including both public and private school students, there are strikingly few children at high levels of performance.

2. Second, less than a third of the advantaged, socio-economic elite, children enrolled in public schools reach the SDG minimal target of PISA Level 2 or higher.

3. Third, there are, on average, important disadvantages in learning outcomes along five characteristics: gender, rurality, home language, being an immigrant, and socio-economic status. However, given the massive global inequalities in learning, remediating only these within-country inequalities in learning without improving the effectiveness of the education system, while undoubtedly important for equity and justice, leads to only modest gains towards the SDG targets.

The SDG goals, targets, and indicators for education raise an important empirical question. How much of the deficit from achieving the goal of universal minimum proficiency in the desired skills/competencies/capabilities in reading and mathematics is because a country's education system is not “inclusive and equitable” and how much is because its education system is not effective in providing “quality education”? This question informs ongoing debates about how much of the education agenda should focus on improving the overall effectiveness of teaching and learning practices in the education system so that all children learn more versus how much effort should go into the equalization in education outcomes across identified categories of potential disadvantage. This paper uses the recently available PISA for Development (PISA-D) data from seven countries—countries that this paper shows have learning levels typical of developing countries, to examine the relative contributions of overall country performance and five commonly identified social elements of education disadvantage (sex, rurality, native language, immigrant status, and household socioeconomic status) to the observed learning outcomes of students enrolled in public schools. This allows an empirical discussion of the relative contribution of “country” versus “within-country inequalities” to achieving the SDG targets.

The empirical answer is very clear: inclusion is very far from enough. This paper shows that even if the inequalities in enrollment and learning across all of the five measured student characteristics (girl/boy, rural/urban, native/migrant, native speaker of instruction language/other language in the home, SES) were completely eliminated by raising education outcomes such that every child had the same learning achievement at age 15 of the children enrolled in public school who were advantaged and socioeconomically elite, the vast majority of children in every PISA-D country (except Ecuador) would still be far from getting a globally minimally adequate education. When the public schools are providing even the social and economic elite an education that does not reach the global minimum—which are levels of performance lower than demonstrated to be achievable by even other very poor countries—the only possible remedy is system-wide change in the quality of education. In the existing learning conditions of the PISA-D countries, policies and programs that are tailored exclusively to remediating socioeconomic gaps can only be a modest part of the progress needed to achieve the SDGs.

At the core of every education system are the classroom and the school—and the teaching and learning practices enacted there. To address the learning crisis and prepare their youth with at least the minimal skills/competencies/capabilities needed by successful adults, countries need teaching and learning practices in their classrooms and schools that are both effective and inclusive. High-performing education systems, having achieved high levels of effective practices—hence where the vast majority of students are already above global minimums in learning foundational skills but in which social inequalities remain—often shift their focus to inclusivity. The challenge facing education systems where even minimum learning goals are not being achieved broadly (or, worse, even achieved by the society's advantaged, socio-economic elite) is more severe. These systems need to simultaneously: (a) expand enrollments and grade attainments so that every child completes basic schooling, (b) dramatically improve the effectiveness of teaching and learning practices in basic schooling so that those who complete it are actually equipped with the learning, skills, and competencies they need to thrive as adults, and (c) improve the inclusivity of teaching and learning practices so that children who begin school at a disadvantage (e.g., first-generation learners, from poorer households, girls, non-native speakers of the instruction language) have those disadvantages ameliorated rather than exacerbated by the schooling system. Getting to universal mastery of adequate skills, capabilities, and competencies will require education systems that do (a) (universal completion of basic schooling) and (b) (effective instruction) and (c) (inclusion).

This paper examines nationally representative PISA-D data for seven countries that have learning profiles typical of developing countries and compares them to OECD and benchmark education systems participating in the PISA study. The study documents that even in the sampled universe of students (i.e., both public and private school students), there are strikingly few children at high levels of performance. Second, the analysis examines the learning profiles of the socioeconomically advantaged children and finds that less than a third of those enrolled in public schools reach the SDG minimal target of PISA Level 2 or higher. Third, the disadvantages in learning outcomes are examined along five characteristics: gender, rurality, home language, being an immigrant and socio-economic status are important. A counter-factual analysis is carried out to measure the progress towards the SDGs after closing the within-country inequalities related to the five measures of disadvantage. It is found that given the massive global inequalities, only remediating these within-country inequalities in learning without improving the effectiveness of the education system as a whole, while undoubtedly important for equity and justice, leads to only modest gains towards the SDG targets. The study also provides a simulation and discusses to what extent targeting that is based on student characteristics in low-performing countries can be ineffective.

The study makes three main contributions to the existing literature. First, it relies on nationally representative data for countries with typical characteristics of developing countries and examines learning outcomes of the advantaged and the elites in these countries. It focuses on both the learning-performance elite and the learning performance of the socioeconomic elite. Second, it provides estimates of the stock of the learning-performance elite, and presents the learning profiles of the socioeconomically advantaged elite. Third, it presents an analysis examining all the typical measures of disadvantage simultaneously (i.e., gender, rural location, migration status, language spoken at home, and socio-economic status) rather than a single dimension or cross-tabulations, and carries out counterfactual analysis that shows that closing all the socioeconomic gaps by raising disadvantaged groups would not be enough to meet the SDGs.

Section 2 briefly reviews the existing literature, presents the background information on PISA for Development, a description of the dataset and the analytical sample, and descriptive statistics. Section 3 describes the empirical strategy and method used in the paper and discusses the main findings of that analysis. Section 4 discusses the core facts derived from the empirical analysis and presents the results related to the SDG progress from equalization and inclusion, and discusses the limitations of targeting disadvantaged groups in low-performance countries. Section 5 concludes and discusses policy implications.

2. Literature Review and Description of PISA-D (Background, Results, and the Size of the Learning-Performance Elite)

This section provides a brief review of the relevant existing studies, background information on the PISA for Development, a description of the dataset and analytical sample, as well as the core descriptive statistics, and the estimation of the size of the learning performance elite.

2.1. Literature Review

This paper contributes to two main areas of the existing literature. First, the study focuses on the PISA-D data and complements and contributes to the literature that has examined variation in learning outcomes across countries and inequalities based on the international achievement studies like PISA and Trends in International Mathematics and Science Study (TIMSS) and other regional assessments. The paper briefly reviews here a subset of these studies. Filmer and Pritchett (2009); Hanushek and Woessmann (2011) and Woessmann (2016) provide a comprehensive review of this area of the literature, which has examined how cross-country variation in learning outcomes is related to the international variation in schooling institutions and governance. This area of research has also highlighted the cost of low educational achievement and the gains associated with improvements in the skills of a nation's labor force (e.g., Hanushek and Woessmann 2010). More closely related to the present paper are the studies that examine the role of family background in explaining variation in student achievement. Schuetz, Ursprung, and Woessmann (2008) construct a measure to assess how students’ educational performance is related to family background by relying on a sample of 54 countries from the TIMSS study. Freeman and Viarengo (2014) examine the role of family background and parental involvement in affecting learning outcomes by relying on the mostly middle-income countries and advanced economies in the PISA data. In another study, Anand et al. (2019) examine the contribution of households and schools to the variation in learning outcomes by focusing on three low-income East African countries. In contrast, Gust, Hanushek, and Woessmann (2022) by relying on raw micro data for 159 countries estimate the present value of lost world economic output related to missing the foundational goal of universal basic skills to be equal to US$700 trillion over the remaining century.

Second, the present analysis with the PISA-D data adds to the growing empirical literature from a variety of countries suggesting, that grade-attainment expansion and inclusivity without effective instruction, are not viable plans for achieving the learning goals of the SDGs or eliminating learning poverty. Crouch and Rolleston (2017), Crouch and Gustafsson (2018), and Crouch, Rolleston, and Gustafsson (2020) review the available evidence from a variety of sources and argue that, while there are important inequalities by gender or wealth, the inequalities in learning across country systems are much larger and that progress in average achievement in low-performing systems comes from “bringing up the left tail”—reducing the proportion of students at low levels of learning. Patel and Sandefur (2020) use the results of a “Rosetta Stone” assessment containing questions from multiple different assessments to construct a linkage between a variety of international assessments to put them onto a common scale. They use data from the household questionnaires on household assets to show that in the low-income and lower middle-income countries, the predicted average score of even students from the richest households are below 400 (which is regarded as the “low international benchmark” as defined in the international studies PIRLS and TIMSS, where the norm for OECD countries is constructed to be 500). Pritchett and Sandefur (2017) use the literacy assessment in the Demographic and Health Surveys in 53 countries to show that only about half of adult women in these countries can read a single sentence even after completing six years of primary schooling. Akmal and Pritchett (2021) use the ASER/UWEZO data from five countries (India, Pakistan, Kenya, Tanzania, and Uganda) to show that even if the poor (from households in the bottom 40 percent by an asset index) had the enrollment and learning achievement of children from the richest 20 percent of households, these countries would still be far from achieving even minimal levels of numeracy. Beatty et al. (2018) show that large expansions in the number of youth who completed junior and senior secondary school between 2000 and 2014 did not lead to improvements in the cohort mastery of even grade-school arithmetic. Kaffenberger and Pritchett (2020) use a structural model of learning to show that if learning is increasingly ineffective as children lag behind then efforts to expand enrollment that do not improve teaching and learning practices lead to little or no progress on SDG goals. Rose (2015) presents a discussion of the lessons learned from the EFA Global Monitoring Reports and emphasizes the need for the post-2015 agenda to focus on achieving equity in educational opportunities, as well as monitoring the progress of the disadvantaged groups.

One contribution of the present paper is that it examines all the standard elements of disadvantage (socioeconomic status, sex, residence, immigrant status, language spoken at home (if different from language of assessment)) rather than a single dimension or cross-tabulations. Moreover, this papers value added is in looking at within-country differentials and comparisons of “adjusted” students across countries as opposed to the comparisons of average scores in the existing literature.

2.2. PISA for Development (PISA-D): Measuring Learning Levels for SDG Progress

An approach that has been adopted to monitor the progress on advancing Sustainable Development 4 is to rely on large-scale international student assessments that provide consistent and comparable measures of learning proficiency (e.g., UNESCO 2022). The minimum proficiency level is the threshold of basic knowledge in a specific subject that is measured by the learning assessment in the international studies. In this regard, studies such as PISA and TIMSS have been used to measure the share of students who meet the minimum proficiency levels (in mathematics and reading) at the end of lower secondary education, and PIRLS to measure learning outcomes in primary education.

PISA is an internationally standardized study conducted by the OECD that has measured learning outcomes of in-school 15-year-olds in OECD and partner countries on a 3-year cycle since 2000. PISA is designed to assess students’ skills and their ability to apply knowledge to real-life situations in three domains: reading, mathematics, and science. In each wave of PISA, one of those subjects is the focus domain. The PISA study design aims at national representativeness and cross-country comparability through a national double-level sampling. In the first stage, schools with 15-year-old students at the time of the assessment are sampled from a comprehensive national list of PISA-eligible schools. For the sampled schools, a complete list of 15-year-old students is produced. In the second stage, within schools, students are sampled for the test. The PISA 2015 study (mostly directly comparable to PISA-D) was carried out in the 35 OECD countries and 37 partner countries and economies, covering over 540,000 students.

PISA for Development (PISA-D) was a pilot exercise of extending PISA to low- and middle-income countries launched in 2014 as a response to the demand of the international community for better global data on learning achievement in the context of the United Nations’ Sustainable Development Goals (OECD 2017). PISA-D was implemented within the overall PISA framework and in accordance with PISA technical standards. The PISA-D was deliberately different from PISA in several ways (OECD 2018a): (1) an equal focus on the three test domains; (2) the use of test instruments that allowed for the accurate measurement of student ability at lower levels of proficiency than was possible in PISA; (3) the introduction in the student/household background questionnaires of items more relevant to low-income and middle-income countries; and (4) for a subset of the PISA-D countries at a later stage, a learning assessment for out-of-school children. The learning results for in-school children were released in late 2018.¹

The PISA-D assessment data that was used in this study includes 34,604 students from 7 countries: Ecuador, Guatemala, Honduras, Cambodia, Paraguay, Senegal, and Zambia. The country assessments were carried out in different years: Ecuador (2014), Guatemala (2015), Honduras (2016), Cambodia (2016), Paraguay (2015), Senegal (2015), and Zambia (2014). While there are only seven PISA-D countries reporting results and they are not meant to be “representative” of the developing world, comparing their levels of development and their learning levels to other measures of learning (e.g., DHS reading results, World Bank Human Capital index² data for learning, Patel and Sandefur (2020)) suggests these countries are typical among lower-income developing countries. Nothing in the present paper should be seen as singling out these particular countries for criticism for their low levels of learning as they are typical in learning performance among countries at their level of development. Instead, these countries should be praised for their willingness, courage, commitment, and capability in participating in the PISA process and in allowing the results and raw data to be disseminated.

PISA defines levels of proficiency characterized by the competencies students demonstrate at those levels as ranges of PISA scores. Table 1 shows the numerical ranges for the PISA levels and the descriptions of the skills and capabilities that students at those levels would display in the assessment. The PISA assessment was normed so that the mean student in the OECD was at 500 (within Level 3), and the standard deviation across OECD students was 100.

SDG Goal 4 aspires to “Ensure inclusive and equitable quality education for all. . . .” The various targets bring learning front and center: Target 4.1 refers to “relevant and effective learning outcomes”; Target 4.4 to equipping youth with “relevant skills”; and Target 4.6 refers to “all youth” achieving “literacy and numeracy.” Indicator 4.1.1 measures the proportion of youth “at the end of lower secondary achieving at least a minimum proficiency level in (i) reading and (ii) mathematics.” It has been agreed that the PISA Level 2 proficiency “marks the baseline level of proficiency at which students begin to demonstrate the competencies that will enable them to participate effectively and productively in life as continuing students, workers and citizens” (OECD 2017).³

Level 2 (above 420.7 but below 488.38 for mathematics) includes very basic skills like interpretations that “require no more than a direct inference” and “can extract relevant information from a single source” “employ basic algorithms” and make “literal” interpretations. Level 1c (up to a score of 295.47) is at a level that is roughly a “rote” level of understanding in which students can only respond adequately to the simplest of questions when presented in a familiar format.

This study's use the PISA assessment does not imply that cognitive skills are regarded as the only important outcome of education systems nor that PISA is an ideal and perfect measure of the entire variety of skills youth might need in the twenty-first century nor even that PISA is a perfect measure of cognitive skills. But the PISA assessments are a reliable and valid measure of at least one set of skills, a set of skills that all education systems include in their stated goals and curricula, and the PISA results are an accepted measure for the SDGs.

2.3. PISA-D Results

The results for the seven countries for the three learning domains with average score, percent below Level 2, and percent at Level 1c or below are presented in fig. 1. One can see that the differences in performance between the PISA-D countries, Vietnam (chosen as a very high learning-outcome performer among poorer countries), and Denmark (chosen as a typical OECD country) are massive. Overall learning results for the PISA-D countries show very low learning levels, with between 70 and 95 percent of those 15-year-olds enrolled in school not reaching PISA Level 2 in each of the learning domains (except for Ecuador).⁴

The large proportion of students in PISA-D countries at Level 1c or below is why the PISA instrument had to be extended to accurately measure learning differences within the subcategories of Level 1. For instance, in Denmark in mathematics only 0.4 percent of students (less than 1 in 200) were at Level 1c or compared with more than 25 percent (1 in 4) students in all PISA-D countries, except Ecuador, at this rudimentary level. Whereas only 13.6 percent of students were not meeting the SDG in mathematics in Denmark, fewer than 13.6 percent reached the math SDG in all countries, but Ecuador and (barely) Honduras. In the lowest-performing of the PISA-D countries (Cambodia, Senegal, Zambia), more than 90 percent of enrolled students were not reaching PISA Level 2 proficiency in each of the three subjects.

2.4. The “Learning Performance Elite” in PISA-D Countries Is Tiny

So far this study has focused on the implications for achievement of minimum levels of learning of the relatively low average performance of the PISA-D countries. Another objective of every country's system of basic education is to produce students with high cognitive skills, who are prepared to go on for advanced academic and professional training. This study takes PISA Level 4 and above as the threshold for the global “learning performance elite.” This is not a very demanding standard as roughly 30 percent of OECD students reach Level 4 or above, and the description of the level emphasizes the “limited” range of skills and the requirement that students can reason in “straightforward contexts.”

Table 2 shows the estimates of the absolute number of individuals in an age cohort who reach Level 4 or above in mathematics, reading, and science. In mathematics it is less than 1,000 individuals for all countries but Ecuador. This ranges from as low as 5 in Zambia to only 519 in Honduras. In the sciences there are similarly few, ranging from zero (that is, no one who took the assessment was estimated to be at Level 4 or above in Cambodia) to 271 in Honduras. In reading there are fewer than 1,300 in every country but Ecuador, ranging from only 4 in Cambodia and 5 in Zambia to only 1,276 in Guatemala.

When one thinks of the range of needs in a modernizing economy for people with skills that require math and science (e.g., engineers of all types (civil, mechanical, electrical), doctors, nurses, chemists, statisticians, actuaries, computer and IT specialists, agronomists, accountants, even perhaps some economists, etc.) it is sobering to think that in any given cohort that is leaving basic schooling and perhaps looking forward to tertiary (college and university training) there are, in a country like Guatemala of 390,000 people in an age cohort, only 141 in mathematics and 177 in science who at age 15 who are well prepared at even a modest global standard for further study. Compare that with Vietnam, that has 240,605 students in mathematics and 281,245 students in science at Level 4 or above.

Perhaps even more fundamentally, the ability to read, process, understand, and analyze texts is central to an entire range of professions and occupations, including those in math and the sciences, and having less than 1,500 students a year emerging from their basic education with reading at level 4 or above poses severe limitations to development prospects.

Table 2 makes explicit and translates into absolute numbers of youth, rather than percentages of a distribution, what is present and implicit in other global comparisons. The tiny numbers of youth above a modest global threshold is a necessary consequence of four known facts about many (but, as with Vietnam, not all) developing countries: (1) the central tendency (mean, median) learning performance is very low, (2) the variability of learning performance across students (e.g., standard deviation, 90th–10th percentile range) is not absolutely larger than in high-performing countries, (3) the tails of the distribution of learning thin out (even if not exactly Gaussian Normal), and (4) many countries have modest-sized total populations.

Two important implications.

One, an examination of the production of global performers in the PISA-D countries reveals that these numbers are extremely low compared to the typical OECD country, or to a high-performing developing country like Vietnam. Existing research has documented to what extent it is the level of cognitive skills of the workforce that plays a fundamental role in affecting its competitiveness and economic growth (Hanushek and Woessmann 2011). To be competitive in a global knowledge-based economy, these countries will need to focus on not only the expansion of opportunity generally but also on the upper tail (Pritchett and Viarengo 2009).

Two, as a preview of the results in the next sections, the tiny size of the global learning-performance elite itself implies that most of the children of the national elites in these countries, however privileged they are relative to others in their own country, are not getting a globally good education. Take Guatemala, where undoubtedly the social and economic inequalities are massive. In mathematics and science less than one-tenth of 1 percent of enrolled 15-year-olds are reaching the globally modest threshold of PISA Level 4 performance. Hence even if every one of the “global learning performance elite” were from the SES elite of the “1 percent” this would still imply only 1 in 10 of the SES elite children are getting a globally adequate education.

3. The Poor Performance of the Advantaged and Elite Students

Calculating the potential progress towards SDG learning goals from greater inclusion and eliminating disadvantages requires: (1) measures of learning outcomes, (2) measures of advantage and socioeconomic status (SES), (3) an estimated empirical model that links the two. This section presents the available measures of advantage and socio-economic status, a descriptive model, and then the results of an estimated empirical model that links the two.

3.1. Measuring Household Socioeconomic Background and Advantage

This study uses the PISA Economic, Social and Cultural Status (ESCS) index, which summarizes family/household conditions that are considered to be relevant to the child's education. The student questionnaire provides information on the highest level of education of either parent, highest occupation status of either parent, and home possessions. Home possessions are measured through the availability of 16 household items including some that are proxies of family income and wealth (e.g., works of art, appliances, cars) and some that are proxies for a household conducive to student learning (e.g., books in the home, a student's own room, availability of computers, educational software, internet). PISA constructs the ESCS from those data using weights from principal component analysis.⁵ The ESCS index is constructed in the same way across the participating countries. In PISA 2015 the index was standardized to have mean 0 and variance 1 for the average student in OECD countries. In PISA-D the index was extended to include other household items typical of low-income and middle-income countries.

This ESCS index both is and is not comparable across countries. Mechanically, the index is comparable in that the same weights were used on the same variables in each country to construct the index. But there are two important cautions in comparing the index internationally. One, those who took the test in the developing countries tended to come from the higher economic strata of these countries, as only those in school at age 15 were in the PISA sample. For instance, as shown in table 3, in Zambia only 36 percent of children aged 15 were in the PISA sample. Of those Zambian households in the PISA sample: 28 percent of children had a computer at home that they could use for school work; 51 percent of households had a bank account; 17 percent of students had a father who completed the second stage of tertiary education; and occupationally 9 percent of were professionals and 6 percent managers. This is clearly not nationally representative of Zambian households. In contrast, in Denmark 89 percent of 15-year-olds were in the sample and hence the ESCS is necessarily closer to being representative of the population. A child that is 2 standard deviations higher in the ESCS distribution of those in the PISA sample is at an ESCS score of 1.31 in Zambia and 2.33 in Denmark, but due to selection into schooling the Zambian child in the PISA sample is even more elite relative to the entire population in Zambia than in Denmark.

The second caution is that this index is not based on comparable monetary units and hence does not measure household per capita consumption/income or wealth.⁶ A person with a tertiary degree who is a professional will get the same weight for those characteristics in Zambia or Denmark, whereas their income and assets are likely to be much lower. Hence, when comparing a household with the same ESCS, say a household at 1.31 in Zambia and a household at 1.31 in Denmark the Danish household will have a much, much higher level of material goods or income/consumption. Other issues related to the comparability across countries and over time have been acknowledged in the existing literature.⁷ It is likely that the ESCS index is a reasonable proxy for SES as it relates to education performance within each country, and not terribly implausible for comparing across countries at similar levels of income, but not very comparable between PISA-D countries and high-income countries like Denmark, except in within-country relative terms (e.g., “here is how Danes who are elite relative to other Danes compare to Hondurans who are similarly elite relative to other Hondurans”).

In addition to an indicator of the child's household SES this study includes indicators for sex (male/female), rural/urban residence, speaking the language of instruction in the home, and immigrant status, each of which is, in the cross-tabulations reported in PISA and PISA-D reports and tables, commonly (though not always in all subjects and in all countries) associated with learning performance.

3.2. Empirical Method

The purpose of this study is very different from the much more common use of OLS (and other more sophisticated) regression (and, more recently, experimental) techniques in the enormous “education production function” literature. This literature uses student test scores, student characteristics, and observed variables about the classroom, teacher, school, and school system to investigate the correlates and, aspiration ally, the causes of learning. A typical question in this literature is whether a child would learn more in a larger or smaller classroom or whether a child would learn more with a classroom teacher with this or that academic qualification. In the “education production function” literature, the student characteristics, like gender or ESCS, are treated as “control” variables included to compare, say, class size, for “otherwise identical” students on the “control” variables to parse out the “treatment effect” or causal impact of class size.

The interest of this study is the opposite. It is not using student characteristics as “controls” to estimate a causal impact of something else, rather its interest is in the observed, descriptive, differences in learning by the student characteristics, irrespective of their causal pathways. For instance, this study does not include measures of school resources. Perhaps one reason students in rural schools have lower scores is that rural schools deploy fewer resources per student. But this is not essential for this study's purposes as it is not trying to decompose the “total” OLS conditional mean on “rurality” into its “partial” derivative causal pathways but rather it is interested in the conditional mean of rurality itself. Hence the many concerns about the impossibility of interpreting regression coefficients as causal parameters and about the external validity of the parameters are not relevant.

The study uses OLS just for “descriptive predictions,” which is analogous to cross-tabulations on multiple categories. The PISA-D reports contain tables with cross-tabulations of results by the categories that are being used, and some analysis, like that of Rose (2015), compares cross-tabulations across multiple categories, comparing, say “poor rural girls” to “rich urban boys.” The study uses OLS regressions not only for correlations among the indicators but also allows SES as a continuous variable (allowed to be nonlinear), which makes it possible to examine the performance of the SES “elite” whereas the cross-tabulations approach often forces bunching of SES in categories like quintiles in order to have adequate observations in each cell.

The analysis estimates the following empirical model. For each country/region, regressions are reported where the dependent variable is the estimated score of student i, for each of the subjects s, math, reading, science.⁸

The dependent variables included in each regression are:

• |$Ag{e}_{is}$|⁠, the age of student i measured in years.

• |$Femal{e}_{is}$| is a binary indicator taking value 1 if the student i is female, and 0 otherwise.

• |${R}_{is}$| is = 1 if the student i attends a school located in rural area, and 0 otherwise.

• |${L}_{is}$| is = 1 if the student speaks at home a language different from the one spoken in school, and 0 otherwise.

• |${I}_{is}$| is = 1 if the student is either a first- or second-generation immigrant in the country, and 0 otherwise.

• |$\nabla ESCS$| is a matrix with the linear, squared, and cubic values of the PISA index of the economic, social, and cultural status (ESCS); hence|$\ \gamma $| is a 3 × 1 vector including the three estimates for the coefficients associated with the first-, second-, and third-degree polynomial of the ESCS index included in the matrix |$\nabla ESCS.$|

The sample is limited to students enrolled in public schools from the seven participating countries in PISA-D. It was decided to limit the analysis to public school students in analyzing inequality for two reasons. One, the study is intrinsically interested in comparing the performance of public schools across countries in the world and in the systematic determinants of why public schools perform extremely well in some countries but are weak in others (Pritchett and Viarengo 2015). Two, earlier research examined variations in learning achievement across the private and public schools (Pritchett and Viarengo 2015), and there are massive differences across countries in the assessed learning of students enrolled in public and private schools, and these differences are some complex (and mostly unknown) mix of selectivity of students and causal impacts (Patel and Sandefur 2020, Akmal and Pritchett 2021). The present paper focuses only on the learning inequalities of those who are within the public system and brackets (in this paper) the additional inequalities generated by selection into and (perhaps) additional learning gains in private schools.

In addition to the seven PISA-D countries, regressions were also performed for two selected countries and two variants of the OECD region that participated in PISA 2015. Vietnam was chosen, as it is a developing country with roughly OECD levels of PISA performance. Denmark was chosen as a typical OECD country as its scores are very near the OECD averages.

3.3. Regression Results

The five indicators that are included (sex, rurality, immigrant status, language at home, and socioeconomic status) are typically statistically and practically significant, even conditional on the other factors. That said, there is large heterogeneity on the magnitudes of the specific co-variates (1) across PISA-D countries, (2) across the subjects, and (3) between PISA-D and OECD countries.

Table 4.A reports the regression results for the seven countries in PISA-D (Cambodia, Ecuador, Guatemala, Honduras, Paraguay, Senegal, and Zambia) and table 4.B for the countries/regions that are used for comparison from PISA 2015 (Denmark, Vietnam, OECD, and OECD excluding Mexico (which the analysis takes as the typical result for high-income industrial countries)).⁹

3.3.1. Regression Results for ESCS

ESCS. The ESCS results are the most robust, which should not be surprising, as the ESCS index was constructed to reflect household conditions that were widely believed to be associated with child academic performance. Apart from Senegal, the index of economic, social, and cultural status is positively associated with positive educational outcomes in all countries and in the pooled OECD regressions (with and without Mexico). The analysis allows the ESCS index to enter nonlinearly (a cubic functional form). The nonlinear terms are only rarely statistically significant. While the linear ESCS terms are robustly associated, the magnitude varies substantially across the PISA-D and the comparator countries (but this is potentially confounded by the magnitudes of the other terms).

Figure 2 shows predicted PISA math scores by ESCS level for advantaged (male, urban, nonimmigrant, speaking the assessment language in the home) public school students for the seven PISA-D countries, Vietnam, and Denmark. Overall, the figure shows that advantaged, SES elites students do better than students within their countries with lower SES, but their average performance remains well below the globally minimum standards (i.e., SDG 4, which is below Level 2 in the PISA proficiency scale). Figure 2 illustrates two points.

First, within each country there typically is a large learning advantage to SES. Figure 2 shows for each country the difference in predicted score between students at average ESCS versus average ESCS plus two standard deviations of ESCS. This ranges between only 2 (Senegal) and 52 points (Ecuador), with a median of 33 points (Guatemala). This predicted difference between ESCS elite and average students is smaller in absolute terms in PISA-D than in Vietnam or Denmark, though quite similar in relative terms (as the average level is much higher in those two countries). The PISA-D countries do not have more unequal learning outcomes of public-sector–enrolled students by ESCS than does a typical OECD country or a high-performing developing country like Vietnam.

Second, fig. 2 shows for each country the predicted PISA math score for a child from a household with an ESCS index of 0.62, which is the median across the PISA-D countries of the average ESCS plus two standard deviations. This predicted score ranges from 311 in Senegal to 413 in Ecuador, with a median of 369. Shown for each country is the gap between the country and Vietnam, a country that has similar, but lower, ESCS than the typical PISA-D country. That gap is enormous for all of the countries. For the median PISA-D country, Honduras, the gap is nearly 200 points. Given that the standard deviation among public school students in Honduras is only 70 points this means the equivalent ESCS elite student in Vietnam would be almost three standard deviations ahead of the same (observationally equivalent) child in Honduras. Or, on the crude assumption that the typical gain from a year of schooling is 30 PISA points, a Honduran student at age 15.5 is almost seven grade levels of learning behind a Vietnamese child.

Vietnam has experienced a significant expansion of education over the past 20 years, reaching close to universal completion rate at the primary level, and high enrollment rates at the secondary level. This has happened together with performance on standardized tests that is close to one of the best-performing education systems in advanced economies. Dang et al. (2020) relying on Blinder-Oaxaca decompositions of the gap in average test scores of PISA 2012 and 2015 between Vietnam and the other participating countries find that differences in child, household, and school characteristics do not help much in explaining the learning performance gap, which appears to be a country-wide system impact.

As mentioned above, Patel and Sandefur (2020) create a mathematics score comparable to the TIMSS learning assessment across 65 countries using a “Rosetta Stone” approach to link regional assessments to a common basis. They also estimate for each household in each country the PPP equivalent per capita income. While the purely economic status of the household should not be expected to be strictly comparable to the study's PISA ESCS index results, the broad patterns are similar. They find that for children from households at the same household income (PPP$1,000) the scores range from 300 to 500 in their sample of countries. They also find the score-household income gradients are no higher for lower-income than for higher-income countries. Their figure 11 shows that in low-performing countries even those in the income elite have very low performance and the average in the top reported income category for both low-income and lower-middle-income countries is below 400.

3.3.2. Regression Results for Other Social Characteristics

Age. The analysis includes age but not grade in order to adjust all predicted values to a child of the same age but the grade for age is one pathway whereby learning inequality by characteristics is manifest, and the study did not want to parse that out, and hence older children are likely to score higher from having achieved more years of schooling. In the OECD (less Mexico), the estimate is 13 points, but the age effects show variation among countries. In Cambodia (23.1) and Zambia (17.8), the estimated coefficient is quite large and statistically significant. In contrast, in Guatemala (1.3), Honduras (6.2), and Senegal (10.3), the estimated coefficient is not statistically significant. Because only students between 15.25 and 16.33 years old, at grade 7 or higher are eligible for PISA, age is actually correlated with learning outcomes through grade. In Guatemala, Honduras, and Senegal, a weaker correlation is observed between age and grade than what is observed for the other PISA-D countries (these figures are available from the authors).

Gender. The estimated coefficient for girls differs as well (and the differences across subjects are important, which is discussed below). In “OECD less Mexico” girls score 9.6 points lower on average. In Zambia and Vietnam the coefficient is positive but not statistically significant. In contrast, in five of seven PISA-D countries the coefficient is substantial (>9 points), negative, and statistically significant.

Rural residence. There are also mixed results with respect to rurality, with striking differences between developed and other countries. In the “OECD less Mexico,” the coefficient is almost exactly 0. In contrast, it is negative in each developing economy (both PISA-D and comparator) except for Ecuador and statistically significant in four of the seven PISA-D countries, with magnitudes ranging from −7.1 (Paraguay) to −34.7 (Guatemala).

Speaking another language at home. This is robustly associated with lower results. In the OECD less Mexico the impact is 16.6 points. The estimated coefficient is negative in all PISA-D countries and statistically significant in five of the seven, with magnitudes ranging from −4.1 in Senegal (insignificant) to a huge −74.4 points in Honduras. In the PISA comparator countries it is small and negative in Vietnam and substantial in Denmark. The differences are likely related to the differences in the frequency and social conditions of school versus home language. In Senegal the coefficient is small, but 95 percent of students report speaking another language; in Guatemala the coefficient is substantial, and only 10 percent of the students report speaking another language at home.

Immigrant status. Not surprisingly, given the very different patterns of who is a migrant across the world and their status and conditions vis-à-vis the native residents, the coefficient on immigration status shows wide variation. The “OECD less Mexico” result shows a small coefficient (1.8) with a large standard error (2.9) almost certainly reflecting the heterogeneity across OECD countries, as it is very large and negative in Denmark (−31.5). Being a first- or second-generation immigrant has a large and positive coefficient in Ecuador, Paraguay, and a substantial magnitude and negative coefficient in Cambodia, Guatemala, Honduras, Zambia and Vietnam (and small and insignificant in Senegal).

The analysis now examines the patterns by looking at the differences in the “typical” magnitude of disadvantage across subjects. Table 5 shows the median estimate of learning-score disadvantage for each subject (mathematics, reading, science) for each category of disadvantage (gender, residence, language, and immigrant). The analysis compares the PISA reported estimates of differences both unadjusted and adjusted for student socio-economic status alongside the regression estimates, which shows the comparative results across categories and across subjects by category are robust.

For instance, it is striking that the median gap between boys and girls and urban and rural students is about the same for mathematics (−15.2 and −16.6) but, whereas the learning gap is quite similar across subjects across the urban/rural gap (−16.6, −18.6, −13.0) it is strikingly different across subjects between boys and girls. The gap is large in mathematics whereas in reading girls perform substantially better than boys, +9.2, while the gap is negative for girls but modest in size in science, −5.9. This suggests that the key to understanding the learning gaps between boys and girls lies in specific gendered teaching practices by subject (as the evidence is inconsistent with this being an innate difference between the genders as the gap is zero or positive in favor of girls in mathematics in many countries) whereas the sources of the urban/rural gap seem to be robust across subjects. Another intriguing example is the disadvantage from not speaking the language of assessment in the home. In the median across the countries, this is the largest single disadvantage in learning (and larger than the typical gaps from differences in socioeconomic status). Not surprisingly, this disadvantage is largest in reading, but is also large in mathematics and in science. This is suggestive that the disadvantages from lower language skills spill over into the other subjects.

Figure 3 presents the graphical representation of these results by subject together with the country-specific display of the estimated coefficients, which highlight the significant variation across countries in the estimated learning disadvantage related to the five categories of disadvantage.

4. Advantaged, ESCS Elites of PISA-D Countries Perform Badly by Global Standards: SDG Progress from Inclusion/Equalization Is Modest

This section discusses the learning outcomes of the advantaged elite and carries out a counter-factual analysis to estimate the progress towards SDGs if every child had the same learning profile as the socioeconomically advantaged. It then presents the learning profiles of the advantaged middle class and shows that their learning levels are very low. Finally, it discusses the limitations of targeting learning interventions based on student characteristics in low-performance countries.

4.1. Learning Outcomes of the Advantaged and Socioeconomic Elite

The analysis starts by focusing on the distribution of learning outcomes for the advantaged children in ESCS elite households. By advantaged the study takes the four characteristics measured by PISA-D as common indicators of learning disadvantage and included in the regressions above: gender, rural residence, speaking the language of instruction at home, and being a native of the country. Children are called “advantaged” if they are male, urban residents, natives of the country, and speak the assessment language in their household, even though it was seen above that these characteristics do not convey advantage in each subject in each country. Moreover, the use of the designation of “advantaged” is socially constructed and does not imply that these categories are necessarily sources of advantage, nor that they are normatively justified sources of advantage—this article is certainly not claiming, for instance, that it is everywhere and always an “advantage” to be born male or not be a migrant. By ESCS elite the analysis means those who are at two standard deviations above the mean ESCS for their country. If ESCS had a standard Gaussian Normal distribution these would be households at the 97.5th percentile of the ESCS distribution.¹⁰

Figure 4 shows the predicted score for the advantaged, ESCS elite student for all seven PISA-D countries and all three subjects (mathematics, reading, and science). The advantaged, ESCS elite have low performance in all subjects. Importantly, in (nearly) all countries and in (nearly) all subjects the average of the advantaged, ESCS elite students are below PISA Level 2 proficiency—the only exceptions are Ecuador in reading and science and Paraguay (barely) in reading.

Figure 4 also compares the performance of the advantaged, ESCS elite with similarly situated students in a typical OECD (Denmark) or a high-performing developing country (Vietnam). The results for Denmark and Vietnam are the predicted score of a student in that country that is advantaged and at the same level of ESCS as the average PISA-D country elite. The average PISA-D ESCS plus two standard deviations score is 1.04, about a standard deviation above the OECD average. As shown in fig. 2, in mathematics the advantaged elite students in PISA-D countries are from 100 (Ecuador) to almost 200 (Senegal) points behind those in Denmark, and about 200 points behind Vietnam in the median PISA-D country.

Using the regression results derived from equation (1), where the same explanatory variables are included in the empirical model but the dependent variable is whether a child was above the PISA Level 2 threshold for math, reading, and science, the analysis carries out a counterfactual analysis to address the following questions: “what if every child in the country was both (a) in school at age 15 (so all unenrolled children and drop-outs up to that age were eliminated) and (b) had the same learning outcomes as the assessed advantaged, ESCS elite students?” The tables using estimates of a binary indicator of above PISA level 2 and Probit estimates are available from the authors upon request.

Figure 5 shows the results for all three subjects for the seven PISA-D countries and for Vietnam at the average ESCS for the elite of the PISA-D countries. It is clear that even if the expansion to universal enrollment were achieved and even if all learning inequalities across (public) school children were completely erased (which, as many of these inequalities are present in all countries, is an extremely unlikely scenario), so that every child’s grade completion and learning were raised to have the same level that the advantaged ESCS elite children now enjoy, countries would still be far from achieving the SDG targets. In mathematics, all of the PISA-D countries (except Ecuador) would still have less than a quarter of their 15-year-olds reaching the SDG level of PISA Level 2 proficiency. In reading, the outcomes are more variable, and three countries (Senegal, Cambodia, and Zambia) would also still be less than one-quarter whereas Guatemala and Honduras would be less than one-half. In science, there is roughly the same pattern with Senegal, Cambodia, and Zambia at one quarter or less and Guatemala, Honduras, and Paraguay at less than half. The contrast with Vietnam is striking. In all three subjects, more than 90 percent of children at the PISA-D elite levels of the ESCS are over the SDG threshold. These results put a strong upper bound on what can be achieved from reducing inequality in access alone.

Table 6 shows the details for mathematics and compares the advantaged ESCS elite to other students. The results show two aspects of inequality. Relative to other children in the same country the elite are much more likely to reach PISA Level 2 proficiency: in Guatemala the average student who took the PISA-D assessment only had a 2.5 percent chance whereas the advantaged ESCS elite student has an 18.2 percent chance, which is more than 7 times higher. However, in the global comparison, 18.2 percent implies that only one in five of the advantaged, ESCS elite students in Guatemala are reaching even a global minimum threshold.

This analysis implies that the elimination of socioeconomic differentials in education outcomes in both enrollment/grade attainment and learning can be only one part of any PISA-D country's plan to reach the education SDGs—and in the lower-learning countries an empirically small part. Moreover, the assumption of a complete elimination (or even substantial reduction) in the learning differences by socioeconomic status is extremely implausible (Pritchett and Viarengo 2021). That is, some socio-economic gaps in learning persist in relatively equal countries like Denmark and Finland, which benefit from highly effective education systems and generous and comprehensive social programs (Pritchett and Viarengo 2021). Progress to reach the SDG will require both expansions in grade attainment and the learning of those in school to improve for (nearly) everyone, and by very large amounts.

4.2. Very Low Learning Levels of the Advantaged Middle Class

Crouch and Rolleston (2017) and Crouch, Rolleston, and Gustafsson (2020) emphasize the importance of “raising the floor.” Table 7 uses the PISA-D data on Level 1c or below learning, which, as was seen in table 1 above represents a truly minimal level of learning. The upper threshold for Level 1c for mathematics is about 300, which is 120 points below Level 2. Table 7 shows the percentage of children predicted to be at Level 1c or below in mathematics and reading for the average child and for the advantaged child at average ESCS. Overall, the table shows that even the advantaged “middle class” (mean of ESCS) students often reach only very low levels of learning (Level 1c or below). This shows that in five of the seven PISA-D countries more than a quarter of the male, urban, native, dominant-language-speaking, “middle class” children were at Level 1c or below. Of course this was higher than for children that did not have all these advantages, but not that much higher. In Cambodia, 41 percent of students (across all conditions) were predicted to be at Level 1c or below and 29 percent of the advantaged, average ESCS students who were also at that level.

This examination of very low performance is important because it illustrates the typical (or at least very common) teaching and learning practices in these countries. That is, very low learning levels of rural girls from poor households who are in the first generation to attend school and who do not speak the language of instruction in the home can be understood as the result of a cumulative of disadvantages (see section 3.3.2): rural schools often have difficulties attracting quality teachers, gender bias can work against an emphasis on girl's education, lack of household income leads to multiple causal pathways of disadvantage, children who are part of the first generation to attend school cannot get as much help from parents, lack of instruction in the mother tongue can make it difficult for students to navigate their early school years. But what is striking about the analysis of the frequency and inequality at very low learning levels is that in five of the seven PISA-D countries (i.e., Zambia, Senegal, Paraguay, Cambodia, Honduras) in mathematics at least one in four students (>25 percent) who face none of the four structural disadvantages and is not relatively poor (has average ESCS) nevertheless has very low educational achievement. This examination of very low performance is important because it illustrates the typical (or at least very common) teaching and learning practices in these countries.

As Crouch and Rolleston (2017) emphasize, if one wants to eliminate low levels of learning one needs to focus on eliminating low levels of learning. This might seem a truism, but many would argue one should focus on targeting children with the characteristics that are associated with low learning. However, in low-performing education systems it is not just the “excluded” or “marginalized” who are getting a very weak education. A focus on “universal, early, conceptual and procedural mastery of foundational skills” is a way of “bringing up the bottom tail” of learning which then, necessarily raises the low levels of achievement of the disadvantaged more than others (since, at the margin, they start from lower levels). But this is a systemic focus to bring about global equity by focusing on attaining global levels rather than an exclusive focus on the differentials across categories. It is important to acknowledge that system improvements to “raise the left tail” can arise from explicit programs to teach foundational skills in the local language (e.g., teaching in the local language especially in the first years of compulsory education can help children to acquire core skills such as cognitive skills in numeracy)¹¹ or select more female teachers (e.g., in rural areas especially in remote locations, and at higher levels of education). These are two policies that would necessarily address also rural/urban or gender gaps.

Moreover, recent research has further highlighted the importance of soft skills in affecting future labor market outcomes, and to what extent achievement tests do not adequately measure these skills (e.g., Heckman and Kautz 2012) and has also raised concerns that global learning goals, while important, might not have many cost-effective solutions relative to other policy priorities (e.g., Jackson et al. 2020; Evans and Hares 2021).

4.3. The Analytics of Targeting Learning Interventions By Student Characteristics

The increased attention to income/consumption poverty as an objective of development has led to a very large literature on the targeting of antipoverty efforts in developing-country contexts in which income/consumption was not administratively observable and hence to investigation of a variety of modes of targeting, such as proxy means tests or self-selection or geographic targeting (or combinations). However, with goals other than income/consumption poverty the distinction has to be made between targeting based on observed student characteristics that are associated with student enrollment or student learning that are conditional on enrollment and targeting based on direct measures of enrollment or learning. If the association between education outcomes (enrollment or learning) conditional on enrollment is weak, then the targeting of education efforts based on student characteristics faces three challenges: (1) much of the effort can be inframarginal; (2) the “exclusion error” of those who should get the intervention based on directly measured learning but who are not in the targeted category can be large; and (3) if the goal for learning is very far from current average performance, the gains from targeting relative to the goal can be very small, or negative.

The magnitude of all three of these issues vary massively across countries depending on their average learning performance. While targeting based on student characteristics might be attractive in high-learning-performance countries, the exact same targeting can be ineffective, or even counterproductive, in low-learning-performance countries. And this is especially so when, as has been shown to be the case in tables 4.A and 4.B, the association between student characteristics and learning outcomes of the enrolled is very weak.

To illustrate this point, the analysis performs a simple simulation exercise in which the score of each individual student i in country c is a function of their household SES and a country-specific constant (equation (2)) and a random error term.

In order to roughly mimic the actual PISA and regression results, the standard deviation of the error term is set to 71 (which produces a total standard deviation of scores of 75), the income term is set to 25 (roughly the linear slope from regressions above), the SES is randomly distributed with a standard deviation of 1 (hence in this simulation a student with SES a standard deviation above the mean scores 25 points higher). These parameters are set to be the same for all countries, and then the average performance is determined by varying the country constant from 260 to 510 (to generate the average outcomes from Senegal to Vietnam).

In this simulated data, it is possible to carry out precise “poverty” targeting by allocating a learning intervention to the 20 percent of students with the lowest SES. The analysis assumes that the learning impact of the program is 25 points, which is very large, at one-third of a student standard deviation.

With the simulated outcomes, repeated with 10,000 students for 1000 iterations for each parameter set, with the poverty/SES targeted learning intervention the analysis makes two calculations. First, it calculates the “exclusion error,” which is the fraction of students who were below the learning threshold but who did not benefit from the poverty-targeted learning intervention. Second, it calculates the gain in the percent above the threshold from the poverty/SES targeted-learning intervention versus and random allocation of the intervention across students.

The first point in fig. 6 is that when average performance is very low relative to the learning threshold and hence nearly every student is going to be below the minimum learning threshold, then poverty targeting is going to produce massive (near the maximum) “exclusion error” as nearly all of those below the learning threshold are not “poor.” Even at the average learning outcomes of Honduras (the second-highest performer in mathematics) the exclusion error is 78 percent. This illustrates the point of Crouch and Rolleston (2017) that a strategy of “raising the floor” of learning performance requires targeting learning performance, and this is not at all the same as targeting SES or (even less so) student characteristics.

The second point is that for all of the PISA-D countries a poverty-targeted learning intervention would do less to raise the percentage of students above the SDG threshold than would a randomly assigned intervention. This result is partly an artifact but nevertheless is conceptually important. The result is an artifact in that in low-performing countries the low SES students are so far from the SDG that even a large intervention is going to raise their scores but still not push them over the SDG threshold. Therefore, only when average performance is close to the threshold does the positive impact of SES targeting in making it more likely to reach below-threshold students offset the (perverse) “distance” effect that leads low SES students to be below the threshold even with the intervention.

Although this finding is to some extent a mathematical artifact (as other weights for the learning gains that did not have thresholds would show the same or better results for SES targeting), it does raise an important conceptual point. If the “minimum” learning threshold really is appropriately set at something like PISA Level 2, then one cannot just lower the goal to make SES targeting attractive. That is, there might be a temptation to react to being very far from the “minimum” learning threshold to argue that the goal is unachievable, and hence in order to “effectively” target resources one should lower the threshold and goals for learning. However, it is far from obvious that this is an attractive or effective approach to the political and pragmatic realities of creating education systems that provide effective teaching and learning practices. In this regard, Vietnam, though an outlier, has shown that it is possible to achieve very high learning even in materially straightened circumstances. Moreover, there is no evidence that any of the higher learning gains for the poorer SES students in higher-performing countries is due to lower gradients on SES as both the absolute and relative SES gradients tended to be higher, not lower, in high-performing countries (as documented in Pritchett and Viarengo (2021)).

5. Conclusion

The idea of “inclusion” and “inequality” are overwhelmingly powerful as they tap into deeply rooted concepts of fairness, equity, and justice. But in many developing countries today there does not exist a high-quality public education system for children to be “included” into. In the PISA-D countries, the advantaged (male, urban, native speaker, nonimmigrant) and SES elite children at age 15 do know much more than their compatriot disadvantaged and non-elite (and poor) children—having higher grade attainment and steeper learning profiles. Yet, even the advantaged elites are effectively “excluded” from a publicly provided education of minimum global quality. Public schools with effective teaching and learning practices are just not robustly and routinely available to any students, even the countries’ advantaged elite.

In the six low-performing PISA-D countries¹² even if education were maximally “inclusive” in that (1) all children were in school and (2) all children had the same learning outcomes in school as advantaged, SES elite students: (a) 80 percent or more of children would not reach the SDG minimum learning level of SDG in mathematics, (b) there would still be only a tiny percentage reaching a “global learning performance elite” of PISA Level 4 or above. Moreover, for the advantaged students at median SES more than one in four (except in Guatemala) would be below PISA-D Level 1c, hence still functionally illiterate and innumerate at age 15.

To address the learning crisis (nearly) everyone has to learn much, much more per year of schooling. Countries need teaching and learning practices in their classrooms and schools that are both effective and inclusive. Vietnam demonstrates that it is possible to have essentially OECD levels of learning even at the economic levels of the PISA-D countries—and episodes of large improvements at scale show that countries/regions can accelerate progress (Stern et al. 2021). But progress requires that entire systems of basic education have to change towards greater effectiveness in conveying the skills of literacy and numeracy: inclusion alone can provide only modest gains.

Data Availability

The data underlying the empirical analysis of this paper is publicly available. It can be downloaded from the following OECD websites: PISA-D (data; codebooks and questionnaires) https://www.oecd.org/pisa/pisa-for-development/database/ PISA (data; codebooks and questionnaires) https://www.oecd.org/pisa/data/2015database/.

Footnotes

Notes

Martina Viarengo (corresponding author) is a professor in the Department of International Economics at the Geneva Graduate Institute, Geneva, Switzerland; Fellow at the Center for Economic Policy Research, London, UK; and Associate at the Center for International Development, Harvard University, Cambridge, MA, USA. Her email address is [email protected]. Lant Pritchett is the Research Director of Improving Systems of Education (RISE) at the Blavatnik School of Government, University of Oxford, Oxford, United Kingdom; and Associate at the Center for International Development, Harvard University, Cambridge, MA, USA. His email address is [email protected]. This research study was supported by Oxford Policy Management Limited (OPM). The funding source had no involvement in neither the design nor the writing of the paper.

References